Google has released the primary version of its Frontier Safety Framework, a set of protocols geared toward addressing serious risks that powerful Frontier AI models of the longer term could pose.
The framework defines Critical Capability Levels (CCLs), that are thresholds at which models may present increased risk without additional remediation measures.
We then describe various remedies for models that violate these CCLs. The remedies fall into two most important categories:
- Safety measures – Prevent disclosure of the weights of a model that achieves CCLs
- Deployment Remedies – Preventing abuse of a deployed model that achieves CCLs
The Google framework will likely be released in the identical week OpenAIThe Superalignment security teams disintegrated.
Google appears to be taking potential AI risks seriously, saying: “Our preliminary analyzes of the research and development areas of autonomy, biosecurity, cybersecurity and machine learning.” Our initial research suggests that the powerful capabilities of future models in these areas are most certainly to pose risks represent.”
The CCLs that the framework addresses are:
- autonomy – A model that may expand its capabilities by “autonomously acquiring resources and using them to run and maintain additional copies of itself on rented hardware.”
- Biosecurity – A model that may significantly enable an authority or non-expert to develop known or novel biothreats.
- Internet security – A model able to fully automating cyberattacks or allowing an amateur to perform sophisticated and serious attacks.
- Machine learning R&D – A model that would significantly speed up or automate AI research in a state-of-the-art laboratory.
Of particular concern is the autonomy CCL. We've all seen sci-fi movies where AI takes over, but now Google says future work is required to “guard against the chance of systems acting hostile to humans.”
Google's approach is to periodically review its models against a series of “early warning scores” that flag a model that could be approaching the CCLs.
When a model shows early signs of those critical capabilities, remedial actions can be applied.
An interesting comment within the framework is that Google says: “A model may reach evaluation thresholds before remedial motion at an appropriate level is prepared.”
A model under development could have critical features that may very well be abused, and Google may not yet have a strategy to prevent this. In this case, based on Google, the event of the model can be placed on hold.
Perhaps we will take comfort within the incontrovertible fact that Google appears to be taking AI risks seriously. Are they being too cautious or are the potential risks listed within the framework a cause for concern?
Let's hope we don't discover too late. Google says: “Our goal is to implement this initial framework by early 2025, which we imagine needs to be well before these risks materialize.”
If you might be already concerned about AI risks, Reading the framework will only increase these fears.
The document notes that the framework will “evolve significantly as our understanding of the risks and advantages of frontier models improves” and that “there is important scope for improvement in understanding the risks posed by models in numerous areas.” .
